Jaeyoung-Lim
commented
2 years ago Thanks! Awesome work!
Putting a screenshot of the information visualization on the dataselection for reference
Some comments / issues that I have encountered.
- When I tried using the data selection tool, I ran into the following error. It seems that the residuals calculations are assuming the full dataset while the pipeline is running on a subset that is selected. (Edit: Turns out this is just a bug in master: https://github.com/ethz-asl/data-driven-dynamics/issues/187)
------------------------------------------------------------------------------- Complete results saved to: model_results/multirotor_model_2022-02-16-10-22-50.yaml ------------------------------------------------------------------------------- Traceback (most recent call last): File "Tools/parametric_model/generate_parametric_model.py", line 140, in <module> start_model_estimation(**vars(arg_list)) File "Tools/parametric_model/generate_parametric_model.py", line 122, in start_model_estimation model.compute_residuals() File "/home/jaeyoung/dev/data-driven-dynamics/Tools/parametric_model/src/models/dynamics_model.py", line 488, in compute_residuals y_forces_pred[2::3] = y_pred[int(2*self.y_forces.shape[0]/3):self.y_forces.shape[0]] ValueError: could not broadcast input array from shape (2694,) into shape (14775,) make: *** [Makefile:34: estimate-model] Error 1
Calculation of informational gain for each parameter in dataframe. Can be accessed by saving the frame to csv in the visual data selection screen. Gives great feedback to user which maneuvers gather data for which parameter.
- I didn't quite understand what this means. Could you elaborate? Additionally would it be possible to enable a separate plot for this?
pascalau
This Pr adds averaged FIM which addresses the issue of singular FIM matrices for the minimum eigenvalue metric. The resulting informational plots are much easier for a user to understand and highlights better important sections of the log. Other features include:
Only change that should be thoroughly discussed is the change of the default estimator. I observed that the QPOptimizer sometimes struggles to find correct parameters that allow the vehicle to take off or on the other hand converges always to the same solution even if the log is highly deficient (which might be intentional).